Re: 2.6.26-rc1 regression: ISA DMA broken (bisected)

[Rene Herman <rene.herman@keyaccess.nl>]

On 14-05-08 17:40, Rene Herman wrote:

CC list trimmed as now PNP and ALSA specific.

> Like this? With this (on top of the previous patch setting the
> dma_mask ofcourse) legacy ISA actually appears to be fine but it's
> then ISAPnP which goes bonkers again. Sigh. Getting an allocation
> failure. Don't understand why yet since pnp_alloc_dev() definitely
> sets the mask already. Will stare...

You're in a maze of struct device *s, all alike... I was passing the 
pnp_card->dev instead of the initialized pnp_dev->dev.

And, not doing so brings out a difference between ISAPnP and legacy ISA 
again insofar that legacy ISA does not consist of cards with multiple 
devices. We just have the single struct device * for the ISA device.

This therefore would be the easiest solution (and works fine) but seems 
a bit of a hack. Bjorn, do you have an opinion? If I abstract things out 
a bit more I might be able to do this nicer. One might on the other hand 
argue that the dma_mask is going to be constant for all card devices so 
might as well just use the card dev.

sound/{isa,oss} together with drivers/isdn/hisax/ are the only pnp_card 
users.

diff --git a/drivers/pnp/card.c b/drivers/pnp/card.c
index a762a41..a2842a7 100644
--- a/drivers/pnp/card.c
+++ b/drivers/pnp/card.c
@@ -7,6 +7,7 @@
 #include <linux/module.h>
 #include <linux/ctype.h>
 #include <linux/slab.h>
+#include <linux/dma-mapping.h>
 #include <linux/pnp.h>
 #include "base.h"

@@ -167,6 +168,9 @@ struct pnp_card *pnp_alloc_card(struct pnp_protocol *protocol, int id, char *pnp
 	sprintf(card->dev.bus_id, "%02x:%02x", card->protocol->number,
 		card->number);

+	card->dev.coherent_dma_mask = DMA_24BIT_MASK;
+	card->dev.dma_mask = &card->dev.coherent_dma_mask;
+
 	dev_id = pnp_add_card_id(card, pnpid);
 	if (!dev_id) {
 		kfree(card);